MFAP-Net: A Study on Deep Learning Semantic Segmentation Framework for LiDAR Point Clouds

Yawu Liu, Lifu Chen, Bingnan Wang, Maosheng Xiang, Haoda Chen, Junqi Zhao

Abstract: LiDAR is widely applied in diverse fields such as surveying, mapping, military operations, agriculture, and geology, contributing significantly to social development and technological progress. However, accurately classifying 3D point clouds from LiDAR encounters challenges, including unclear boundaries and insufficient semantic information extraction. This paper introduces the MFAP-Net framework, leveraging deep learning networks. MFAP-Net incorporates MLSE for 3D position coding, 3D coordinate mapping, and 2D localization, capturing spatial characteristics and boundary details. DFAP integrates coordinate information from various dimensions to address semantic boundary challenges effectively. Penalty terms in MLSE-DFAP prevent data redundancy, enhancing semantic segmentation performance by minimizing l2 distance between dimensions. Evaluating the framework using 110Kv, 220Kv, and 500Kv laser point cloud data for power corridors, and comparing with existing networks, MFAP-Net achieves a remarkable average classification accuracy of 95.6% for power lines, towers, buildings, plants, and ground—outperforming other networks. Empirical studies confirm the efficacy of our module.